Every change, in the open.

Start Here now opens with "Your first 10 minutes": four steps to run stage 1, and nothing else to read until it's done. Everything that was there before follows, to read after stage 1. A first-time user said the kit looked like too much to start, so the first screen now asks for one thing.

• No prompt changed, so nothing was re-tested: the v1.5.2 results apply.

The Engine now runs the kit 1.5.2 stages; until now it ran the kit 1.4.1 stages. It picks up everything in v1.5, v1.5.1 and v1.5.2 below.

• New plain-word commands for build sessions: “start a session” and “close the session” (or “we're done for today”). Closing a session rewrites AGENTS.md “Current state” as four lines, adds any lesson to “Lessons learned”, and checks off finished slices in PLAN.md.
• Documents are always saved in your project folder, in the same reply that finishes the stage.
• Tested in Claude Code with the Engine installed as a plugin: 4 scenarios, 3 runs each. The first build passed 7 of 12 runs (it once saved IDEA.md in its own folder, and Claude Code took “close the session” as a request to save its own notes); after those two fixes, 12 of 12.

After a BUILD or SHRINK IT FIRST verdict, stage 1 now ends with one plain sentence naming what's still missing: there's no spec yet, and without one an AI builder fills the gaps with its own guesses. It's skipped when the verdict recommends an existing app or a quick simple build. Only stage 1 changed. The Engine has it from Engine v1.5.0.

• Every stage 1 scenario re-run three times: 3 of 3 each on Sonnet 4.6 and Haiku 4.5.
• Two new scenarios check the new sentence: 3 of 3 on both models, after correcting a test rule that also failed ordinary advice naming third-party tools. Both grades are published on the open testing page.

Not re-tested: neither change alters what the prompts ask for, so the v1.5 results apply.

• Stage 6: one line called the security check “13-item”; it has 14 items, as the rest of the prompt says.
• Start Here: the skills/ line now says what the skills are, three optional helpers Claude loads when they're relevant, instead of “Claude runs the stages without copy-paste”.

Keeping long build sessions on track. A long chat makes the AI follow its instructions less reliably, so the build stage now keeps each chat short and keeps what matters in your project files. Based on Anthropic's “Effective context engineering for AI agents”. Only stage 5 changed. The Engine has them from Engine v1.5.0.

• One chat per slice, with three signs to start a new chat early: the AI forgets a rule in AGENTS.md, asks something you already answered, or undoes earlier work.
• Session close now deletes AGENTS.md “Current state” and rewrites it as four lines, instead of letting it grow every session.
• AGENTS.md has a new “Lessons learned” section: one line for anything a fix taught you, so the next session knows it too.
• A new prompt for Claude Code and Cursor sends a helper agent to look into a question and bring back a short summary, so your main chat stays short.
• The Claude Skills match. Tested: the affected scenarios (T10, E5, E10) and a new Session close test (E13) pass 3 of 3 on Sonnet 4.6. T10, E5 and E13 also pass 3 of 3 on Haiku 4.5; E10 was run on Sonnet only.

Blueprint Engine only. Kit unchanged (at the time, the kit download stayed 1.4.1).

• If a project has an AGENTS.md, the Engine reads it first on every request.

Blueprint Engine only. Kit unchanged (at the time, the kit download stayed 1.4.1).

• In stages without a question round, minor unknowns become stated assumptions so the stage finishes instead of stopping to ask.

Blueprint Engine only. Kit unchanged (at the time, the kit download stayed 1.4.1).

• Starting a project: any details a stage still needs are asked in that stage's first round of questions, in one message, never in a separate "basics" message first.
• No app code before PLAN.md exists, even if you insist; the Engine offers the fastest real path to building instead.
• In regular chat, every finished document ends with a clear "Save this as <filename>" line.

Blueprint Engine only. Kit unchanged (at the time, the kit download stayed 1.4.1).

• Starting a project: when you've clearly described a new idea and there's no code yet, the Engine starts stage 1 right away instead of asking whether anything is built.
• Before asking anything, the Engine reads every Blueprint document you already have and never asks for something they already state.
• "Where am I?" now checks coverage: every feature, role and data item in an earlier document must be covered by the later ones, or the later document is flagged as out of date.
• The security gate instruction now points to every checklist item in the stage 6 prompt instead of a fixed count.

No prompt text changed; the v1.4 test results apply to this version.

• Every stage header and Start Here now say the kit is tested with Claude (Sonnet 4.6 or better); other AI chats are untested until results are published. Previous wording said "Claude or ChatGPT".

The v1.3 run scored 19 of 20, 12 of 12 and, for the first time at three runs each, 2 of 5 on the phone scenarios (18 of 20 and 4 of 5 on Haiku). These changes address what was left; the full suite was re-run on v1.4 and the results are on the proof page.

• Stage 2: one or two features over the V1 cap get cut by the AI, with a one-line reason each, and the spec is written in the same reply; it asks first only when you've insisted every feature is a must-have or there are many more than seven. Therapy and session notes, intake forms, medical history and medications are named as health information.
• Stage 3: when it recommends a home-screen web app over an app-store app, it says where the login token lives in the browser and how it is protected.
• Stage 6: the "backed by RLS or a private bucket means Medium" rule now covers client-supplied display fields such as a sender name.
• Stage 1: Rounds 2 and 3 come in the same reply.
• Claude Skills updated to match.
• Test suite: T22 accepts a protected browser token scheme when the AI recommends the web app; T23 accepts the real-phone slice in or before the first milestone.

The v1.2 re-run scored 19 of 20, 12 of 12 and 19 of 20 on Haiku. Two sentences address the two misses, and the phone-app scenarios join the full three-run suite. Results are on the proof page.

• Stage 6: a file-type or size check that lives only in the browser is Medium at most when the bucket is private, per-user and served through expiring links (one review in three had rated it High). The security-gate skill says the same.
• Stage 1: once you've answered, Round 1 is over however many questions were asked; the AI goes to the reframe and verdict instead of asking more. The architect skill says the same.
• Tests: the five phone-app scenarios (T21 to T25) now run three times each on Sonnet and once on Haiku, like everything else, with inputs rebuilt from the v1.3 prompts.

The v1.1 re-run scored 18 of 20, 9 of 12 and 17 of 20 on Haiku. These changes address what was left; the full suite was re-run on v1.2 and the results are on the proof page.

• Stage 2: when it marks health, children's, education or financial data as sensitive, it now also says plainly that a privacy or legal review is needed before launch, and never files that under "Out of Scope".
• Stage 0 and stage 6: a missing feature or an unhandled edge case is not a security gap; it goes in the spec's "not handled" list, never in the severity table.
• Stage 1: once you've answered Round 1, the AI doesn't ask more questions; it states its assumptions and gives the reframe and the verdict.
• AGENTS.md template and the TruckLine example: before marking a slice done, the AI checks the "Known broken" list and asks you to test or fix anything listed there first.
• Claude Skills: the security-gate skill carries the same severity rule as the prompt.
• Test suite: two rules loosened to match v1.1 behavior (a one-sentence reason for not skipping ahead is enough; an explicit ask is optional when the review proceeds with the missing context noted).

Suite v2 ran every scenario three times on Claude Sonnet 4.6 (14 of 20 passed all three), added 12 harder scenarios (5 of 12), and ran the 20 once on Haiku 4.5 (12 of 20). Every raw answer is on the open testing page. These changes address every miss, and the same scenarios are re-run after them.

• Every prompt: if you ask to skip ahead, the AI declines in two sentences and gets on with the step instead of lecturing. It answers in the language you wrote in.
• Stage 1: the Context Carry now carries a "Sensitive" line, so a health, children's-data or safety warning survives into stage 2.
• Stage 2: knows what an IDEA summary looks like and catches an architecture summary pasted by mistake; proposes the V1 cut itself instead of asking you to choose; uses a working name rather than stopping for one.
• Stage 3: knows what a SPEC summary looks like and names the stage a wrong paste came from.
• Stage 4: when you have no references and no brand, it shows six starting styles and asks you to pick before designing anything.
• Stage 5: a missing DESIGN.md summary now stops the plan; it asks first and writes nothing in the same reply.
• Stage 6 and stage 0: Critical/High only when the exact line and the exact exposure can be named; missing hardening is Medium/Low. The gate never refuses to review code it was given, even if the architecture context is missing.
• AGENTS.md template and the TruckLine example: asking for a deletion no longer counts as confirming it; the AI must say what will be lost and wait for a separate yes, and must ask for the exact error before removing anything.
• Claude Skills: the architect skill now carries the Round 1 structure verbatim, so it no longer opens with a verdict.
• Example README: TruckLine is labelled as the fictional worked example it is.

Also in v1.1, from the phone-app work (present in the stage 3, 5 and 6 prompts during the re-run, so covered by those scenarios, but not yet tested with phone-specific scenarios):

• Start Here now says who Blueprint is for: apps where people log in and their data is saved, on the web, iPhone or Android. It also lists projects it doesn't fit.
• Phone apps are now covered from start to finish:
• Stage 3 decides between an app-store app and a web app on the home screen. It keeps secrets out of the installed app, stores login tokens securely, confirms purchases on the server, covers app store payment rules and push notifications, and lists the developer accounts to set up.
• Stage 5 tests on a real phone from milestone 1 and adds slices for account deletion, the store listing and review.
• Stage 6 adds phone self-checks and a 14th review item for phone apps.
• The launch checklist adds an app store section.
• Claude Skills updated to match.
• Fix from phone test run 1: Stage 3 now says plainly what users give up with a home-screen web app versus an app-store app.

Live-tested on Claude: 8 of 8 difficult-user scenarios passed, and the security gate found 10 of 10 hidden flaws.

• Every prompt now includes a map of all stages and their inputs, so the AI can tell you when you've pasted something meant for a different stage.
• The security review now checks all 13 items in order, one line each, before writing its findings, so no item gets skipped.

Test 2 used a difficult freelancer, a client app with children's data, and a messy app with 10 hidden flaws.

• Every prompt now says which stage it is, stops and asks when a required summary is missing, and explains technical terms in plain English.
• Stage 1: client work is handled by phasing an approved feature list instead of arguing about it, with wording to send the client.
• Stage 2: checks your answers for contradictions before writing the spec, and marks sensitive fields.
• Stage 3: asks where user roles are stored and which file storage is public.
• Stage 0: plain-English risk for every gap, and a privacy review flag for children's, health, education or financial data.
• Stage 6: three new checks (file storage, where roles are stored, sensitive data) and a no-code "public files" test.
• Claude Skills rewritten to match the current kit.
• Fixed a template blank that could trip the missing-input rule.

Test 1 took a non-coding food truck owner through all six stages.

• New stage 0 for apps that are already built.
• Stage 3 adds a plain-English summary and a list of accounts to set up.
• Stage 4 asks about your existing brand.
• Stage 5 adds a prompt to fill in AGENTS.md and requires self-contained build prompts.
• Stage 6 explains how no-code builders can get their code for the review.
• Start Here explains how to open .md files and what to do after a "test before building" verdict.
• Added the complete TruckLine worked example. Filled in the license.

• Restructured into one folder per stage, each with a single prompt.md, and one named document saved per stage.

• First draft: prompts and templates for planning, design, building and security.